Re: Recurring panic

On Wed, Jun 05, 2013 at 01:13:45PM +0300, Konstantin Belousov wrote:
K> On Wed, Jun 05, 2013 at 01:50:43PM +0400, Gleb Smirnoff wrote:
K> > On Wed, Jun 05, 2013 at 10:18:21AM +0200, Ian FREISLICH wrote:
K> > I> I have the following recurring panic on all my heavily network
K> > I> loaded -CURRENT routers.  The current process is always different.
K> > I> 
K> > I> Gleb, can you please chime in with what you've managed to uncover.
K> > 
K> > The panics appear on selfd mutex. The mtx_lock value is a free mutex, but
K> > it has 1 extra bit set:
K> > 
K> > (kgdb) p/x sfp->sf_mtx->mtx_lock
K> > $3 = 0x1000004
K> > 
K> > Rarely (only one panic observed) more than one bit is set:
K> > 
K> > $3 = 0x21000004
K> > 
K> > It is important that selfd mutexes are taken from mtxpool(9), which
K> > is allocated at a early boot stage. Thus, across reboots all possible
K> > sfp->sf_mtx mutexes usually fall into the same virtual memory region.
K> > I'm not sure, but I suppose, they fall into same physical region.
K> > 
K> > This can lead one to idea that RAM in the box has problems. But it
K> > is running ECC memory, and it doesn't experience other random panics.
K> > 
K> > The only special about the box is that it is running pf(4) with huge
K> > ruleset and a lot of traffic. So the pf(4) is the number one suspected,
K> > albeit it isn't closely related to selfds.
K> > 
K> So is the virtual address of the corrupted word same for each panic ?
K> If yes, set up the hw watchpoint in ddb.

Nope, they are different, but close to each other, since live in the
same mtxpool.

-- 
Totus tuus, Glebius.